The HPFOREST Procedure

Binned Search

Specify the PRESELECT=BINNEDSEARCH option to select the variable that produces the rule that has the largest worth. Unlike the PRESELET=LOH and PRESELECT=HOTHORN options, the PRESELECT=BINNEDSEARCH option searches for a splitting rule for each variable. This method preferentially selects nominal variables that have many categories. However, if such variables do not exist or if all the variables have a similar number of categories, then the PRESELECT=BINNEDSEARCH option potentially selects better variables than the other methods.

The PRESELECT=BINNEDSEARCH option usually takes longer than the other methods because searching for a splitting rule is more complex. To make it run faster, interval variables are divided into equally spaced bins just before the search. If an interval variable is deemed best, then PROC HPFOREST uses the unbinned data to search for a rule that uses the selected variable. Binning is done independently in every node. The INTERVALBINS= option specifies the number of bins.